← ClaudeAtlas

checkpointlisted

Diff what a devteam project has actually built against its signed charter and file a dated verdict — goal by goal with the evidence recorded, what diverged and whether anyone decided it, which reversible questions the loop settled without the client, and cost against estimate. Run on a schedule, at milestones, and on demand.
alternative-intelligence-cp/claude-skills · ★ 1 · AI & Automation · score 74
Install: claude install-skill alternative-intelligence-cp/claude-skills
# Checkpointing **A verdict with evidence, not a status update.** This is the mechanism that answers the question the whole pipeline exists for: *is what is being built still the thing that was asked for?* Nobody notices drift from inside a task. Every task can be correct, verified and on time while the sum of them quietly becomes a different product. Only a scheduled diff against the charter catches that, and only if it is adversarial (P-32): the question is **"what would have to be true for this to have gone wrong, and is it?"** — never "does this look about right", which it always will. ## 1. Does what exists satisfy the charter? Walk **every charter goal** down the chain and record what you find at each step. Do not skip a goal because you remember it being handled. ``` G-n → the R-n that satisfy it → the tasks that discharged those → the acceptance evidence ACTUALLY RECORDED in their reports ``` **A goal you cannot walk down that chain is the finding**, and it does not matter how confident anyone is that the work was done. "Discharged" means the acceptance criterion was run and its output is in the record (P-5). ```bash python3 "${CLAUDE_PLUGIN_ROOT}/scripts/check_trace.py" . python3 "${CLAUDE_PLUGIN_ROOT}/scripts/check_refs.py" . ``` Then go past what the scripts can see: **read the acceptance criteria and ask whether what was built actually satisfies them**, rather than whether a command exited zero. A test can pass and still not test the requirem